AI Security Risk

An AI Security Risk is a cybersecurity risk that specifically threatens AI system integrity, AI model confidentiality, or AI service availability.

• AKA: AI System Security Threat, Machine Learning Security Risk, AI Vulnerability, Model Security Risk.
• Context:
  • It can typically exploit AI-Specific Vulnerabilitys absent in traditional software systems.
  • It can typically target Training Pipelines, model weights, or inference endpoints.
  • It can typically enable Capability Theft, model manipulation, or service disruptions.
  • It can typically require Specialized Defenses beyond conventional cybersecuritys.
  • It can often involve Supply Chain Attacks on ML frameworks and datasets.
  • It can often motivate Secure Enclaves and confidential computings.
  • It can often cost millions in damages through IP theft or misuses.
  • It can range from being a Data-Level AI Security Risk to being a Model-Level AI Security Risk, depending on its attack surface.
  • It can range from being a Training-Time AI Security Risk to being an Inference-Time AI Security Risk, depending on its attack timing.
  • It can range from being a External AI Security Risk to being an Insider AI Security Risk, depending on its threat origin.
  • It can range from being a Passive AI Security Risk to being an Active AI Security Risk, depending on its exploitation requirement.
  • ...
• Example:
  • Model Attack Risks, such as:
    • AI Model Weights Theft stealing trained parameters.
    • Model Poisoning Attack corrupting training datas.
    • Adversarial Example Attack fooling classification systems.
  • Infrastructure Risks, such as:
    • Training Cluster Compromise accessing compute resources.
    • API Endpoint Exploitation abusing inference services.
    • Dataset Contamination injecting malicious samples.
  • Supply Chain Risks, such as:
    • Backdoored Framework hiding trojan functionalitys.
    • Compromised Pretrained Model containing hidden behaviors.
  • ...
• Counter-Example:
  • Traditional Software Vulnerability, which lacks AI-specific elements.
  • Physical Security Risk, which threatens hardware not models.
  • Operational Risk, which involves process failure not security breachs.
  • Compliance Risk, which concerns regulation not technical threats.
• See: Cybersecurity Risk, AI Safety Risk, Model Security, Adversarial Machine Learning, AI Model Weights Theft, Data Poisoning, Model Extraction Attack, Secure ML, Confidential Computing, AI Supply Chain Security.

---